Foils are insanely cool.

Paymaster support in Starknet React! What I love about this is that it's 1 line change to add the provider and then just switch the send transaction hook.

I appreciate how Gemini pushes back when my initial assumption is wrong. Together with the tasteful response format it's now my favourite model for research.

I got a warning for extremely hot weather today. It's going to be 31C 🤔

Feeling FOMO about cannes. Keep the pictures coming!

What if we called it the RAPID stack? Rust, Arrow, Parquet, Iceberg, DataFusion. /cc Andrew Lamb

underrated aspect of vibe coding: the amount of effort required for a quick hack and proper implementation is almost the same, so may as well do things properly

zed "follow agent" button is good, seeing it jump around files is so satisfying like watching a pro starcraft player

Articles about LLM coding miss that previously writing code and thinking about the problem had to run in serial, now they run in parallel. While the agent is coding, I'm already thinking about the next iteration.

My advice to get Claude to write good Rust: - Implement one module at the time - Reject changes and update the prompt if the solution sucks, telling it to not do what it just did - Add concrete examples to the rules file if Claude makes the same sloppy mistake multiple times

Very good read. I've also found that priming a new session with the content of key files is super helpful to get good results. Never tried ultrathink, usually I don't leave much room for choice

TIL Claude is smart enough to generate docs with `cargo docs` and then grep their content.

Claude generating tests that build a struct and then assert the fields values is the dumbest thing possible.

I'm happy to pay $20/mo for Zed to use an editor that doesn't lag when I work

How do you get claude to stop writing useless comments describing what the next line of code does? I told it to stop commenting WHAT a line of code does and it worked for a while but now it's back to the old habit

Code imports should be at the bottom. I suspect it's much easier to LLMs that way.

Love how fast the model is, but it struggles with editing files. Anthropic cooked with Claude always generating valid changes

For some reason it's now searching the web for documentation. We're back to the dark ages

Kimi K2 endorsed

Using Kimi K2 all morning and it's so good. Doesn't have the annoying personality of Claude and works just as good for me.